29 June 2023 in Scientific Reports
The aim of this study was to develop a convolutional neural network (CNN) for classifying positron emission tomography (PET) images of patients with and without head and neck squamous cell carcinoma (HNSCC) and other types of head and neck cancer. A PET/magnetic resonance imaging scan with 18F-fluorodeoxyglucose (18F-FDG) was performed for 200 head and neck cancer patients, 182 of which were diagnosed with HNSCC, and the location of cancer tumors was marked to the images with a binary mask by a medical doctor. The models were trained and tested with five-fold cross-validation with the primary data set of 1990 2D images obtained by dividing the original 3D images of 178 HNSCC patients into transaxial slices and with an additional test set with 238 images from the patients with head and neck cancer other than HNSCC. A shallow and a deep CNN were built by using the U-Net architecture for classifying the data into two groups based on whether an image contains cancer or not. The impact of data augmentation on the performance of the two CNNs was also considered. According to our results, the best model for this task in terms of area under receiver operator characteristic curve (AUC) is a deep augmented model with a median AUC of 85.1%. The four models had highest sensitivity for HNSCC tumors on the root of the tongue (median sensitivities of 83.3–97.7%), in fossa piriformis (80.2–93.3%), and in the oral cavity (70.4–81.7%). Despite the fact that the models were trained with only HNSCC data, they had also very good sensitivity for detecting follicular and papillary carcinoma of thyroid gland and mucoepidermoid carcinoma of the parotid gland (91.7–100%).
7 September 2023 in Journal of Medical and Biological Engineering
Purpose: Head and neck squamous cell carcinoma (HNSCC) is one of the most common cancer types globally. Due to the complex anatomy of the region, diagnosis and treatment is challenging. Early diagnosis and treatment are important, because advanced and recurrent HNSCC have a poor prognosis. Robust and precise tools are needed to help diagnose HNSCC reliably in its early stages. The aim of this study was to assess the applicability of a convolutional neural network in detecting and auto-delineating HNSCC from PET-MRI data. Methods: 2D U-net models were trained and tested on PET, MRI, PET-MRI and augmented PET-MRI data from 44 patients diagnosed with HNSCC. The scans were taken 12 weeks after chemoradiation therapy with a curative intention. A proportion of the patients had follow-up scans which were included in this study as well, giving a total of 62 PET-MRI scans. The scans yielded a total of 178 PET-MRI slices with cancer. A corresponding number of negative slices were chosen randomly yielding a total of 356 slices. The data was divided into training, validation and test sets (n = 247, n = 43 and n = 66 respectively). Dice score was used to evaluate the segmentation accuracy. In addition, the classification capabilities of the models were assessed. Results: When true positive segmentations were considered, the mean Dice scores for the test set were 0.79, 0.84 and 0.87 for PET, PET-MRI and augmented PET-MRI, respectively. Classification accuracies were 0.62, 0.71 and 0.65 for PET, PET-MRI and augmented PET-MRI, respectively. The MRI based model did not yield segmentation results. A statistically significant difference was found between the PET-MRI and PET models (p = 0.008). Conclusion: Automatic segmentation of HNSCC from the PET-MRI data with 2D U-nets was shown to give sufficiently accurate segmentations.
18 June 2024 in International Journal of Data Science and Analytics
Our aim is to compare different thresholds for a convolutional neural network (CNN) designed for binary classification of medical images. We consider six different thresholds, including the default threshold of 0.5, Youden’s threshold, the point on the ROC curve closest to the point (0,1), the threshold of equal sensitivity and specificity, and two sensitivity-weighted thresholds. We test these thresholds on the predictions of a CNN with InceptionV3 architecture computed from five datasets consisting of medical images of different modalities related to either cancer or lung infections. The classifications of each threshold are evaluated by considering their accuracy, sensitivity, specificity, F1 score, and net benefit. According to our results, the best thresholds are Youden’s threshold, the point on the ROC curve closest to the point (0,1), and the threshold of equal sensitivity and specificity, all of which work significantly better than the default threshold in terms of accuracy and F1 score. If higher values of sensitivity are desired, one of the two sensitivity-weighted could be of interest.
25 September 2025 in Clinical Physiology and Functional Imaging
Purpose: To develop and evaluate a three‐dimensional convolutional neural network for automated classification of PET/MRI images in head and neck cancer (HNC) patients, assessing its performance against radiologist interpretation and its potential as a diagnostic aid.Methods: Data from 202 patients with HNC who underwent 18F‐FDG PET/MRI were used to train and validate PET‐, MRI‐, and PET/MRI‐based models. Of these data, 101 patients were labelled as positive in terms of having HNC, and 101 patients as negative. An additional test set of 20 patients was also evaluated, where 10 patients were labelled as positive and 10 as negative. The model performance was assessed using sensitivity, specificity, accuracy, and AUC. Grad‐CAM was utilised to improve interpretability and classification results on the test set were compared with a radiologist.Results: The PET‐based model achieved an AUC of 0.92 on the test set, with an accuracy of 90%, a sensitivity of 100% and a specificity of 80%. PET/MRI and MRI‐based models underperformed relative to the PET‐based model. The radiologist achieved perfect classification accuracy. Analysis of Grad‐CAM showed that the model classifications are based on real areas of interest. In addition, it gave valuable insight into using similar systems in identifying false positive findings.Conclusion: The PET‐based model demonstrated high sensitivity, indicating its potential as a pre‐screening tool for HNC. However, specificity requires improvement to reduce false‐positive rates. Enhanced datasets and refinement of model architecture will be crucial before clinical adoption. Grad‐CAM provides valuable insights into model decisions, aiding clinical integration.
28 August 2024 in Network Modeling Analysis in Health Informatics and Bioinformatics
A convolutional neural network (CNN) can be used to perform fully automatic tumor segmentation from the positron emission tomography (PET) images of head and neck cancer patients but the predictions often contain false positive segmentation caused by the high concentration of the tracer substance in the human brain. A potential solution would be a one-click annotation in which a user points the location of the tumor by clicking the image. This information can then be given either directly to a CNN or an algorithm that fixes its predictions. In this article, we compare the fully automatic segmentation to four semi-automatic approaches by using 962 transaxial slices collected from the PET images of 100 head and neck cancer patients. According to our results, a semi-automatic segmentation method with information about the center of the tumor performs the best with a median Dice score of 0.708.